The inference of velocity fields from the displacement of objects and/or fields visible within a series of consecutive images over known time intervals has been explored extensively within experimental fluid dynamics. Real image sequences of environmental hydrodynamic flows, however, pose additional challenges for velocity field inference due to factors such as lighting inhomogeneity, particle density, camera orientation and stability. Here we investigate the performance of classical and deep learning based velocity estimation methods on three experimental datasets; a hydrodynamics laboratory dataset of different flow types and two open-source datasets of aerial river footage from field campaigns. The river datasets are accompanied by observational datasets of in-situ measurements. In particular, we investigate the generalisation of deep learning based methods from ideal training conditions to real images. We consider three deep learning approaches; recurrent all-pairs-field transforms (RAFT), a physics-informed approach and an unsupervised learning approach (UnLiteFlowNet-PIV). Results indicate that RAFT, which achieves state-of-the-art performance on particle image datasets, showed good generalisation to the laboratory dataset and field imagery. The physics-informed approach performed similarly to RAFT across the laboratory dataset whilst generalisation to drone-based data proved challenging. Across the laboratory dataset, UnLiteFlowNet-PIV showed good performance within wake regions but an underestimation of channel flows and freestream regions with limited vorticity, also suffering under poor seeding density. Limited fine-tuning of UnLiteFlowNet-PIV on laboratory data, however, led to improved performance in these regions, indicating the potential of the unsupervised learning approach for environmental flows where 2D ground truth data sources are unavailable for training.

中文翻译：

---

从 PIV 到 LSPIV：利用深度学习进行环境流速测量


在实验流体动力学中，已经广泛探讨了从已知时间间隔内一系列连续图像中可见的物体和/或场的位移推断速度场。然而，由于照明不均匀性、粒子密度、相机方向和稳定性等因素，环境流体动力学流的真实图像序列对速度场推断提出了额外的挑战。在这里，我们研究了基于经典和深度学习的速度估计方法在三个实验数据集上的性能;一个不同流量类型的流体动力学实验室数据集和两个来自现场活动的航拍河流镜头的开源数据集。河流数据集附有原位测量的观测数据集。特别是，我们研究了基于深度学习的方法从理想训练条件到真实图像的泛化。我们考虑三种深度学习方法;递归全对场变换 （RAFT），一种物理知情方法和无监督学习方法 （UnLiteFlowNet-PIV）。结果表明，RAFT 在粒子图像数据集上实现了最先进的性能，对实验室数据集和现场图像表现出良好的泛化性。物理信息方法在整个实验室数据集中的表现与 RAFT 相似，而对基于无人机的数据的泛化被证明具有挑战性。在整个实验室数据集中，UnLiteFlowNet-PIV 在尾流区域内表现出良好的性能，但低估了涡度有限的通道流和自由流区域，在种子密度较差的情况下也受到影响。 然而，对实验室数据进行有限的微调导致了这些区域的性能改进，这表明无监督学习方法在 2D 地面实况数据源无法进行训练的环境流中具有潜力。